0 2003 IEEE
0-7803-8185-8/03/$17.00 970
Originalhage Bit-Plane
Decomp o6iti cm
In Fig. 2, white represents a one and black a zero. Both (3) If a secret block is less complex than the threshold ao,
squares, or patches, have the same number of ones and ze- conjugate it to make it more complex. Here the process
ros, hut very different complexities. This shows that one called conjugation, which guarantees that any secret
contains much more visual information than the other. The data can be embedded, is the exclusive OR operation
complex patch (a) has very little visually informative infor- with a checkerboard pattern. The relation a * = 1 - a
mation, therefore it can he replaced with secret data and have holds bue[2], where a and a* are the complexity of a
a very small effect on the image's quality. However, if the given image and that of the conjugated image, respec-
more visually informative patch (b) was replaced, it would tively.
cause noise-like distortion of the definite edges and shapes.
(4) Replace each noise-like block in the hit-planes with a
block of secret data. If the block is conjugated, then
record this fact in a conjugation map.
(5) Also embed the conjugation map in the same way as the
secret blocks.
97 1
3. 3-D SPIHT Video Coding of compression, which is called progressive compression, is
particularly suitable for Internet communication.
3-D SPIHT for video coding was proposed extended from
2-D SPIHT for image coding. 3-D SPIHT has the follow- 4. BPCS Steganography Integrated with 3-D
ing characteristics: (1) partial ordering by magnitude of 3- SPIHT Video Compression
D wavelet transformed video, (2) ordered bit-plane coding,
and (3) exploitation of self-similarity across spatio-temporal
In the 3-D SPIHT video compression, the wavelet coef-
orientation trees (See Fig. 3). Spatio-temporal orientation
ficients of a video are quantized into a bit-plane structure
trees are groups of 3-D wavelet coefficients organized into
and therefore BPCS steganography can be applied in the
trees rooted in the lowest frequency (coarsest scale) subband
wavelet domain. The wavelet coefficients have many image-
with offspring in several generations along the same spatio-
like properties, and BPCS steganography is ideal for exploit-
temporal orientation in the higher frequency subbands. Fig. 3
ing them. The main properties leveraged for BPCS steganog-
shows parent-offspring relationships in the case of two-level
raphy are:
3-D wavelet decomposition. In the spatia-temporal orien-
tation trees, each node has no offspring or eight offspring,
.Correspondence: Spatial areas in each section of the sub-
bands correspond directly to areas in the original image.
which form a group of 2 x 2 x 2 adjacent pixels. The trees
were introduced to exploit self-similarity and magnitude-
. Complexity: The bit-planes at corresponding signifi-
cance levels of the wavelet coefficients and the original image
localization properties in a 3-D DWT video. Typically in
are usually proportionally complex.
case magnitude of a wavelet coefficient in a node does not
. Resilience: Changes in the values of the wavelet coef-
exceed a given threshold, it is very possible that none of its
ficients do not create disproportionately large changes in the
descendants will exceed that threshold.
reconstructed image.
The procedure for data embedding and extraction in 3-D
SPIHT-BPCS Steganography is shown in Fig. 4. 'Ibe entire
process to embed data in 3-D SPIHT-BPCS steganography
follows the solid line arrows in Fig. 4. After 3-D DWT is ap-
plied to an original video, 3-D SPIHT encoder is applied to
the wavelet coefficients and a bit-stream (compressed video
file) is produced. Then the bit-stream is decoded by 3-D
SPIHT decoder and quantized wavelet coefficients are de-
rived'. Using these quantized wavelet coefficients, bit-planes
for the wavelet coefficients can be constructed and used to
embed secret data by BPCS steganography (See the upper
box of the right part in Fig. 4). The quantized wavelet co-
efficients modified by embedding are then subjected to 3-D
SPIHT encoding again to produce a secret-data-embedded
Figure 3. Spatio-temporal orientation tree for bit-stream. The symbol (E) in Fig. 4 depicts that secret
the twdevel3-D DWT. data is embedded. Data embedding in an already compressed
video file is also possible. In this case, the process starts with
The successive approximation method used by the 3-D a compressed video file, i.e., a bit-stream from the bottom of
the middle part in Fig. 4 and follows the same process as the
SPIHT algorithm encodes the wavelet coefficients one bit-
plane at a time, starting with the most significant bit. In 3-D aforementionedone.
SPMT compression, each wavelet coefficient tu is expressed The data extraction procedure follows the dashed arrows
as in Fig. 4. 3-D SPIHT decoding of secret-data-embedded
bit-stream produces secret-data-embedded quantized wavelet
'uf = T(ao + a12-1 +. ' . + a,_12-n+'), a, E {O, 1) ( 2 ) coefficients. Extraction of secret data is carried out by the
BPCS method using the quantized wavelet coefficients. We
where T = 2L'0g2w"-=l (utmaz is the maximum ahsolute assume that the data extraction starts after the entire file of
value among all wavelet coefficients in a 3-D DWT video). the bit-stream has been received.
Since (ao+ulZ-'+. . .+a,-12-"+') is a binary expression,
the 3-D DWT video can be considered to have a bit-plane 5. Experimental Results
structure. 3-D SPIHTencoding is conducted from higher bit-
planes to lower ones. That is to say, encoding starts with The 3-D SPIHT-BF'CS steganography algorithm was im-
more important information, so that decoding can be per- plemented and tested on two standard videos: "Claire" and
formed on the most important information first. Therefore,
even if decoding is discontinued before the end of the video 31n principle the two step of 3-D SPIHT encoding and 3-D SPIHT de-
coding are unnecessary to obtain the quantized wavelet coefficients. How-
file, almost optimal decoding results can still be achieved ever, the two steps are performed so that the bit-stream may be truncated to
under the reduced amount of information. Thus, this type meet pmembedding compression m e requirements.
972
4 1 4
3-D DWT BPCS
(Extraction) (Embedding)
I
wavelet caefficients q m w ' wavelet
I
quantiledwavelef
"Diskus". They consists of 32 frames, each of which is 8-bpp Acknowledgement This work was partly supported by
gray image and 256 x 256 pixels in size. A four-level 3-D the International CommunicationsFoundation.
wavelet transform with the Daubechies 9/7 filter was applied
to videos. The number of hit-planes in the 3-D SPIHT com- 7. References
pression was set to 11 and 12. Here 4 x 4 patch size was
used as an embedding unit and random binary data was used S. Katzenbeisser and F.A.P. Petitcolas, "Information
as secret data. The complexity threshold a. for embedding Hiding Techniques for Steganography and Digital Wa-
was set to 0.3. termarking", Artech House, 2ooO.
Table 1 shows results of embedding experiments where
degradation in video quality has not been perceived after em- M. Niimi, H. Noda and E. Kawaguchi, "A steganogra-
bedding, and compression results without embedding for ref- phy based on region segmentation by using complex-
erence. The PSNR value in the table is the mean for total 32 ity measure", Trans. of E K E , Vol. J81-D-It, pp.1132-
frames. The least significant hit-plane and the two least sig- 1140,1998.
nificant hit-planes were used to embed data for the number E. Kawaguchi and R. 0. Eason, "Principle and appli-
of hit-planes 11 and 12, respectively. The average emhed- cations of BPCS-steganography", Proc. of SPIE, Vol.
ding rate ((embedded data size)/(compressedvideo file size)) 3528, pp.464-473, 1998.
for two videos was around 18% for 11 bit-planes, and 28%
for 12 bit-planes. Fig. 5 shows frames for "Claire" (upper J. Spaulding, H. Noda, M.N. Shirazi and E. Kawaguchi,
row) and "Diskus" (lower row) whose PSNRs after embed- "BPCS steganography using EZW lossy compressed
ding are the lowest among 32 frames. In Fig. 5 , (a) and (d) images", Pattern Recognition Letters, Vo1.23,pp.1579-
are original frames, (b) and (e) are compressed frames with- 1587,2002.
out embedding,and (c)and (f) are embedded and compressed
frames. The PSNRs for (c) and (0 are 43.3dB and 39.4dB, H. Noda, J. Spaulding, M.N. Shirazi, M. Niimi and
respectively. E. Kawaguchi, "Bit-plane decomposition steganogra-
phy combined with PEG2000 compression", Lecture
6. Conclusions Notes in Computer Science, Vo1.2578 (Edited by Fa-
bien A.P. Petitcolas), pp.295-309, 2003.
This paper presented a large capacity steganography
method applicable to compressed video, which is invented
B.J. Kim, X. Zixiang and W.A. Pearlman, "Low bit-
based on BPCS steganography and 3-D SPIHT video com- rate scalable video coding with 3-D set partitioning in
pression. The proposed 3-D SPIHT-BPCS steganography hierarchical trees (3-D SPIHTP: IEEE Trans. Circuits
achieved embedding rates of around 28% of the compressed and Systems for Video Technology, Vol.lO, pp.1374-
video size for twelve bit representation of wavelet coeffi- 1387,2000.
cients with no noticeable degradation in video quality. A. Said and W.A. Pearlman, "A new, fast, and effi-
We should note that the 3-D SPIHT-BPCS scheme is not cient image codec based on set partitioning in hierarchi-
robust with respect to lossy compression in the sense that cal trees", IEEE Trans. Circuits and Systems for Video
lossy compression of an already embedded and compressed Technology, V01.6, pp.243-250, 1996.
dummy video can still easily destroy the embedded informa-
tion. Works are under progress to realize another viderrbased
steganography applicable to Motion-JPEG2000 compressed
video.
973
video # bit-planes # bit-planesused embedded data compressedfile PSNR
for embedding size (bytes) size (bytes) (dB)
11 118175 47.1
Claire 11 1 2 1802 123678 44.4
12 249508 49.7
12 2 69658 264709 45.3
11 316656 44.2
Diskus 11 1 58424 326881 41.7
12 540452 48.8
12 2 173430 567980 41.6
(d) (e) (0
Figure 5. Examples of experimental results: (a),(d) original frames, (b),(e) 3-D SPIHT compressed
frames, (c),(f) embedded and 3-D SPIHT compressed frames.
974